1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@generated from testing/testing_zhemv_mgpu.cpp, normal z -> d, Wed Jan 22 14:40:21 2025
@author Mark Gates
*/
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
// includes, project
#include "flops.h"
#include "magma_v2.h"
#include "magma_lapack.h"
#include "magma_operators.h"
#include "testings.h"
// --------------------
int main(int argc, char **argv)
{
#ifdef MAGMA_HAVE_OPENCL
#define dA(i_, j_) dA, ((i_) + (j_)*ldda)
#define dX(i_) dX, ((i_))
#define dY(i_) dY, ((i_))
#else
#define dA(i_, j_) (dA + (i_) + (j_)*ldda)
#define dX(i_) (dX + (i_))
#define dY(i_) (dY + (i_))
#endif
TESTING_CHECK( magma_init() );
magma_print_environment();
real_Double_t gflops, cpu_time=0, cpu_perf=0, gpu_time, gpu_perf, mgpu_time, mgpu_perf, dev_time, dev_perf;
double Ynorm, error=0, error2=0, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t n_local[MagmaMaxGPUs];
magma_int_t N, Noffset, lda, ldda, blocks, lhwork, ldwork, matsize, vecsize;
magma_int_t incx = 1;
double alpha = MAGMA_D_MAKE( 1.5, -2.3 );
double beta = MAGMA_D_MAKE( -0.6, 0.8 );
double *A, *X, *Y, *Ylapack, *Ydev, *Ymagma, *Ymagma1, *hwork;
magmaDouble_ptr dA, dX, dY;
magmaDouble_ptr d_lA[MagmaMaxGPUs], dwork[MagmaMaxGPUs];
magma_device_t dev;
magma_queue_t queues[MagmaMaxGPUs];
int status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
double tol = opts.tolerance * lapackf77_dlamch("E");
magma_int_t nb = 64; // required by magmablas_dsymv_mgpu implementation
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_create( dev, &queues[dev] );
}
// currently, tests all offsets in the offsets array;
// comment out loop below to test a specific offset.
magma_int_t offset = opts.offset;
magma_int_t offsets[] = { 0, 1, 31, 32, 33, 63, 64, 65, 100, 200 };
magma_int_t noffsets = sizeof(offsets) / sizeof(*offsets);
printf("%% uplo = %s, ngpu %lld, block size = %lld, offset %lld\n",
lapack_uplo_const(opts.uplo), (long long) opts.ngpu, (long long) nb, (long long) offset );
printf( "%% BLAS %s MAGMA 1 GPU MAGMA MGPU Error rel Error rel\n"
"%% N offset Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) to %s to LAPACK\n"
"%%==================================================================================================================\n", g_platform_str, g_platform_str );
for( int itest = 0; itest < opts.ntest; ++itest ) {
// comment out these two lines & end of loop to test a specific offset
for( int ioffset=0; ioffset < noffsets; ioffset += 1 ) {
offset = offsets[ioffset];
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
Noffset = N + offset;
lda = Noffset;
ldda = magma_roundup( Noffset, opts.align ); // multiple of 32 by default
matsize = Noffset*ldda;
vecsize = (Noffset-1)*incx + 1;
gflops = FLOPS_DSYMV( N ) / 1e9;
blocks = magma_ceildiv( N + (offset % nb), nb );
lhwork = N*opts.ngpu;
ldwork = ldda*(blocks + 1);
TESTING_CHECK( magma_dmalloc_cpu( &A, matsize ));
TESTING_CHECK( magma_dmalloc_cpu( &Y, vecsize ));
TESTING_CHECK( magma_dmalloc_cpu( &Ydev, vecsize ));
TESTING_CHECK( magma_dmalloc_cpu( &Ymagma, vecsize ));
TESTING_CHECK( magma_dmalloc_cpu( &Ymagma1, vecsize ));
TESTING_CHECK( magma_dmalloc_cpu( &Ylapack, vecsize ));
TESTING_CHECK( magma_dmalloc_pinned( &X, vecsize ));
TESTING_CHECK( magma_dmalloc_pinned( &hwork, lhwork ));
magma_setdevice( opts.device );
TESTING_CHECK( magma_dmalloc( &dA, matsize ));
TESTING_CHECK( magma_dmalloc( &dX, vecsize ));
TESTING_CHECK( magma_dmalloc( &dY, vecsize ));
// TODO make magma_dmalloc_bcyclic helper function?
for( dev=0; dev < opts.ngpu; dev++ ) {
n_local[dev] = ((Noffset/nb)/opts.ngpu)*nb;
if (dev < (Noffset/nb) % opts.ngpu)
n_local[dev] += nb;
else if (dev == (Noffset/nb) % opts.ngpu)
n_local[dev] += Noffset % nb;
magma_setdevice( dev );
TESTING_CHECK( magma_dmalloc( &d_lA[dev], ldda*n_local[dev] ));
TESTING_CHECK( magma_dmalloc( &dwork[dev], ldwork ));
}
//////////////////////////////////////////////////////////////////////////
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &matsize, A );
magma_dmake_symmetric( Noffset, A, lda );
lapackf77_dlarnv( &ione, ISEED, &vecsize, X );
lapackf77_dlarnv( &ione, ISEED, &vecsize, Y );
/* =====================================================================
Performs operation using cuBLAS / clBLAS
=================================================================== */
magma_setdevice( opts.device );
magma_dsetmatrix( Noffset, Noffset, A, lda, dA, ldda, opts.queue );
magma_dsetvector( Noffset, X, incx, dX, incx, opts.queue );
magma_dsetvector( Noffset, Y, incx, dY, incx, opts.queue );
dev_time = magma_sync_wtime(0);
magma_dsymv( opts.uplo, N,
alpha, dA(offset, offset), ldda,
dX(offset), incx,
beta, dY(offset), incx, opts.queue );
dev_time = magma_sync_wtime(0) - dev_time;
dev_perf = gflops / dev_time;
magma_dgetvector( Noffset, dY, incx, Ydev, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (1 GPU)
=================================================================== */
magma_setdevice( opts.device );
magma_dsetvector( Noffset, Y, incx, dY, incx, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_dsymv_work( opts.uplo, N,
alpha, dA(offset, offset), ldda,
dX(offset), incx,
beta, dY(offset), incx, dwork[ opts.device ], ldwork,
opts.queue );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
magma_dgetvector( Noffset, dY, incx, Ymagma1, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (multi-GPU)
=================================================================== */
magma_dsetmatrix_1D_col_bcyclic( opts.ngpu, Noffset, Noffset, nb, A, lda, d_lA, ldda, queues );
blasf77_dcopy( &Noffset, Y, &incx, Ymagma, &incx );
// workspaces do NOT need to be zero -- set to NAN to prove
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
magmablas_dlaset( MagmaFull, ldwork, 1, MAGMA_D_NAN, MAGMA_D_NAN, dwork[dev], ldwork, opts.queue );
}
lapackf77_dlaset( "Full", &lhwork, &ione, &MAGMA_D_NAN, &MAGMA_D_NAN, hwork, &lhwork );
mgpu_time = magma_sync_wtime(0);
magma_int_t info;
info = magmablas_dsymv_mgpu(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_dsymv_mgpu returned error %lld: %s.\n",
(long long) info, magma_strerror( info ));
}
info = magmablas_dsymv_mgpu_sync(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_dsymv_sync returned error %lld: %s.\n",
(long long) info, magma_strerror( info ));
}
mgpu_time = magma_sync_wtime(0) - mgpu_time;
mgpu_perf = gflops / mgpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
blasf77_dcopy( &Noffset, Y, &incx, Ylapack, &incx );
cpu_time = magma_wtime();
blasf77_dsymv( lapack_uplo_const(opts.uplo), &N,
&alpha, A + offset + offset*lda, &lda,
X + offset, &incx,
&beta, Ylapack + offset, &incx );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Compute the Difference LAPACK vs. Magma
=================================================================== */
Ynorm = lapackf77_dlange( "F", &Noffset, &ione, Ylapack, &Noffset, work );
blasf77_daxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ylapack, &incx );
error2 = lapackf77_dlange( "F", &Noffset, &ione, Ylapack, &Noffset, work ) / Ynorm;
}
/* =====================================================================
Compute the Difference cuBLAS vs. Magma
=================================================================== */
Ynorm = lapackf77_dlange( "F", &Noffset, &ione, Ydev, &Noffset, work );
blasf77_daxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ydev, &incx );
error = lapackf77_dlange( "F", &Noffset, &ione, Ydev, &Noffset, work ) / Ynorm;
bool okay = (error < tol && error2 < tol);
status += ! okay;
if ( opts.lapack ) {
printf( "%5lld %5lld %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(long long) N, (long long) offset,
cpu_perf, cpu_time*1000.,
dev_perf, dev_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, error2, (okay ? "ok" : "failed") );
}
else {
printf( "%5lld %5lld --- ( --- ) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e --- %s\n",
(long long) N, (long long) offset,
dev_perf, dev_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, (okay ? "ok" : "failed") );
}
/* Free Memory */
magma_free_cpu( A );
magma_free_cpu( Y );
magma_free_cpu( Ydev );
magma_free_cpu( Ymagma );
magma_free_cpu( Ymagma1 );
magma_free_cpu( Ylapack );
magma_free_pinned( X );
magma_free_pinned( hwork );
magma_setdevice( opts.device );
magma_free( dA );
magma_free( dX );
magma_free( dY );
for( dev=0; dev < opts.ngpu; dev++ ) {
magma_setdevice( dev );
magma_free( d_lA[dev] );
magma_free( dwork[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
// comment out these two lines line & top of loop test a specific offset
} // end for ioffset
printf( "\n" );
}
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_destroy( queues[dev] );
}
opts.cleanup();
TESTING_CHECK( magma_finalize() );
return status;
}
|
